A BCI System for Imagined Speech Classification Based on Optimization Theory

Abstract

Electroencephalograms (EEGs) are used for establishing a connection between the human brain and the outside environment, so they are widely used in the brain computer interface (BCI). Nowadays, the imagined speech (IS) is a highly promising paradigm of the BCI. It can be used for controlling the external devices directly. However, the features for performing the IS are unknown. Hence, the numerous features are extracted. As a result, the dimension of the feature vectors is extremely large. To reduce the required computation, the clustering is required to be performed in the low dimensional space. Under this circumstance, the transform matrix affects both the dimensional reduction part and the clustering part. In fact, finding the transform matrix and the clustering centers under this scenario is challenging. To tackle this difficulty, this paper provides a modified joint principal component analysis (PCA) and k means algorithm for performing the IS. Here, the interclass separation among the feature vectors is also taken into an account of the problem formulation. In particular, the problem is formulated as a nonconvex constrained optimization problem. The total two norm reconstruction error of the feature vectors as well as the total two norm differences between the feature vectors and the clustering centers in the low dimensional space and the total two norm differences among the clustering centers are minimized subject to the orthogonality of the transform matrix. The numerical computer simulations are conducted based on the multi-class IS classification database. The obtained results show that our proposed method outperforms the various states of the art methods in terms of the clustering accuracy and the average required execution time. Overall, using the BCI system for performing the imagined speech classification plays an important role in the consumer electronics area.